Гібридна згорткова мережа для обробки зображень рентгенівських знімків для виявлення захворювання COVID-19

• Datum: 2022
• Hauptverfasser: Федорченко , Є. М., Олійник, А. О., Степаненко, О. О., Федорончак, Т. В., Чорнобук, М. О., Корнієнко , С. К.
• Format: Artikel
• Sprache: Ukrainian
• Veröffentlicht: Інститут проблем реєстрації інформації НАН України 2022
• Schlagworte: Keras library; decision tree; support vector machine; classification; neural network; perception
• Online Zugang: http://drsp.ipri.kiev.ua/article/view/274945
• Назва журналу: Data Recording, Storage & Processing

The study is devoted to researching the solutions used to detect COVID-19 from X-rays of the lungs. The investigation analyzes various models of machine learning, in particular decision trees, the method of support vectors, neural networks, including convolutional neural networks.  Models based on decision trees, the support vector method, and simple neural networks were built using the Accord.NET library for the C# language. The proposed model based on a convolutional neural network is built using the Keras library for the Python language, which is an extension of another library — Tensor Flow. Based on the review, a decision was made to develop a model based on a hybrid convolutional neural network. For training and testing the model, a publicly available dataset of X-ray images of patients' lungs was used, consisting of images belonging to three classes: lesions of COVID-19, normal state of the lungs, other diseases. Two other datasets were generated from the public dataset by random sampling: a smaller dataset of 450 images and a larger dataset of 1500 images.  As a result, a hybrid convolutional neural network was developed, which achieved a classification accuracy of about 70 % on a smaller data set, and 87 % on a larger one. On an extended dataset of 10,500 images, the model achieved 91 % accuracy. Thus, the proposed model outperformed other considered machine learning algorithms in terms of classification accuracy. This model can be used as an auxiliary diagnostic tool for medical personnel, which will lead to a decrease in the probability of medication errors.